import os import gradio as gr import json import logging import torch from PIL import Image import spaces from diffusers import DiffusionPipeline, FlowMatchEulerDiscreteScheduler from huggingface_hub import hf_hub_download, HfFileSystem, ModelCard, snapshot_download import copy import random import time import re import math import numpy as np import traceback from prompt_rewrite import rewrite import hashlib ################################### base_model = "Qwen/Qwen-Image" # Lightning LoRA info (no global state) LIGHTNING_LORA_REPO = "lightx2v/Qwen-Image-Lightning" LIGHTNING_LORA_WEIGHT = "Qwen-Image-Lightning-4steps-V2.0.safetensors" LIGHTNING8_LORA_WEIGHT = "Qwen-Image-Lightning-8steps-V2.0.safetensors" LIGHTNING_FP8_4STEPS_LORA_WEIGHT = "Qwen-Image-fp8-e4m3fn-Lightning-4steps-V1.0-bf16.safetensors" ################################### def apply_aspect_ratio(ratio): sizes = { "1:1": (1024, 1024), "16:9": (1365, 768), "9:16": (768, 1365), "3:2": (1254, 836), "2:3": (836, 1254), "3:1": (1774, 591), "2:1": (1448, 724), } return sizes.get(ratio, (1024, 1024)) DEFAULT_ASPECT_RATIO = "16:9" # ✅ NUEVO: importar optimización avanzada tipo Qwen-Image-MultipleAngles #from optimization import optimize_pipeline_ LORAS_CACHE = { "data": [], "last_hash": None, } def load_loras_hot(): """Load loras.json and detect changes.""" path = hf_hub_download( repo_id="lichorosario/qwen-image-lora-dlc-v3", filename="loras.json", repo_type="space", ) with open(path, "r", encoding="utf-8") as f: raw = f.read() current_hash = hashlib.sha256(raw.encode("utf-8")).hexdigest() if current_hash != LORAS_CACHE["last_hash"]: LORAS_CACHE["data"] = json.loads(raw) LORAS_CACHE["last_hash"] = current_hash print("🔁 LoRA config updated") return LORAS_CACHE["data"] # Load LoRAs from JSON file def load_loras_from_file(): """Load LoRA configurations from external JSON file.""" try: with open('loras.json', 'r', encoding='utf-8') as f: return json.load(f) except FileNotFoundError: print("Warning: loras.json file not found. Using empty list.") return [] except json.JSONDecodeError as e: print(f"Error parsing loras.json: {e}") return [] # Load the LoRAs #loras = load_loras_from_file() loras = load_loras_hot() # Initialize the base model dtype = torch.bfloat16 device = "cuda" if torch.cuda.is_available() else "cpu" # Scheduler configuration from the Qwen-Image-Lightning repository scheduler_config = { "base_image_seq_len": 256, "base_shift": math.log(3), "invert_sigmas": False, "max_image_seq_len": 8192, "max_shift": math.log(3), "num_train_timesteps": 1000, "shift": 1.0, "shift_terminal": None, "stochastic_sampling": False, "time_shift_type": "exponential", "use_beta_sigmas": False, "use_dynamic_shifting": True, "use_exponential_sigmas": False, "use_karras_sigmas": False, } scheduler = FlowMatchEulerDiscreteScheduler.from_config(scheduler_config) pipe = DiffusionPipeline.from_pretrained( base_model, scheduler=scheduler, torch_dtype=dtype ).to(device) """ # ✅ NUEVO BLOQUE: aplicar AOT optimization (igual que Qwen-Image-MultipleAngles) try: example_args = ( "a cute cat in a spacesuit", ) example_kwargs = dict( num_inference_steps=4, true_cfg_scale=3.5, width=1024, height=1024, num_images_per_prompt=1, ) optimize_pipeline_(pipe, *example_args, **example_kwargs) print("✅ Transformer AOT optimization complete.") except Exception as e: print(f"⚠️ AOT optimization skipped: {e}") """ MAX_SEED = np.iinfo(np.int32).max ### MODIFICACIÓN 1: AÑADIR FUNCIONES PARA GESTIONAR EL HISTORIAL ### def update_history(new_images, history): """Añade las nuevas imágenes generadas al principio de la lista del historial.""" if history is None: history = [] if new_images is not None and len(new_images) > 0: updated_history = new_images + history return updated_history[:24] return history def clear_history(): """Devuelve una lista vacía para limpiar la galería de historial.""" return [] ### FIN DE LA MODIFICACIÓN 1 ### class calculateDuration: def __init__(self, activity_name=""): self.activity_name = activity_name def __enter__(self): self.start_time = time.time() return self def __exit__(self, exc_type, exc_value, traceback): self.end_time = time.time() self.elapsed_time = self.end_time - self.start_time if self.activity_name: print(f"Elapsed time for {self.activity_name}: {self.elapsed_time:.6f} seconds") else: print(f"Elapsed time: {self.elapsed_time:.6f} seconds") def update_selection(evt: gr.SelectData, width, height): selected_lora = loras[evt.index] new_placeholder = f"Type a prompt for {selected_lora['title']}" lora_repo = selected_lora["repo"] updated_text = f"### Selected: [{lora_repo}](https://huggingface.co/{lora_repo}) ✨" examples_list = [] try: model_card = ModelCard.load(lora_repo) widget_data = model_card.data.get("widget", []) if widget_data and len(widget_data) > 0: for example in widget_data[:4]: if "output" in example and "url" in example["output"]: image_url = f"https://huggingface.co/{lora_repo}/resolve/main/{example['output']['url']}" prompt_text = example.get("text", "") examples_list.append([prompt_text]) except Exception as e: print(f"Could not load model card for {lora_repo}: {e}") return ( gr.update(placeholder=new_placeholder), updated_text, evt.index, width, height, gr.update(interactive=True) ) def handle_speed_mode(speed_mode): """Update UI based on speed/quality toggle.""" if speed_mode == "light 4": return gr.update(value="Light mode (4 steps) selected"), 4, 1.0 elif speed_mode == "light 4 fp8": return gr.update(value="Light mode (4 steps fp8) selected"), 4, 1.0 elif speed_mode == "light 8": return gr.update(value="Light mode (8 steps) selected"), 8, 1.0 elif speed_mode == "Wuli-art": return gr.update(value="Light mode (4 steps) Wuli-art selected"), 4, 1.0 else: return gr.update(value="Normal quality (45 steps) selected"), 45, 3.5 @spaces.GPU(duration=70) def generate_image( prompt_mash, steps, seed, cfg_scale, width, height, lora_scale, negative_prompt="", num_images=1, prompt_enhance=False, ): pipe.to("cuda") # if negative_prompt == '': # negative_prompt = "低分辨率,低画质,肢体畸形,手指畸形,画面过饱和,蜡像感,人脸无细节,过度光滑,画面具有AI感。构图混乱。文字模糊,扭曲。" if prompt_enhance: with calculateDuration("Enjancing prompt"): print(f"Calling pipeline with prompt: '{prompt_mash}'") prompt_mash = rewrite(prompt_mash) seeds = [seed + (i * 100) for i in range(num_images)] generators = [torch.Generator(device="cuda").manual_seed(s) for s in seeds] images = [] with calculateDuration("Generating images (sequential)"): for i in range(num_images): current_seed = seed + (i * 100) generator = torch.Generator(device="cuda").manual_seed(current_seed) result = pipe( prompt=prompt_mash, negative_prompt=negative_prompt, num_inference_steps=steps, true_cfg_scale=cfg_scale, width=width, height=height, num_images_per_prompt=1, generator=generator, ) images.append((result.images[0], current_seed)) return images def generate_images_for_prompts( prompts, negative_prompt, steps, seed, cfg_scale, width, height, quantity, # ✅ FIX: ahora entra como parámetro prompt_enhance=False, ): pipe.to("cuda") # if negative_prompt == '': # negative_prompt = "低分辨率,低画质,肢体畸形,手指畸形,画面过饱和,蜡像感,人脸无细节,过度光滑,画面具有AI感。构图混乱。文字模糊,扭曲。" images = [] for prompt in prompts: current_seed = seed if prompt_enhance: prompt = rewrite(prompt) # ✅ FIX: quantity ya no es el componente global; es un int real for _ in range(int(quantity)): generator = torch.Generator(device="cuda").manual_seed(current_seed) result = pipe( prompt=prompt, negative_prompt=negative_prompt, num_inference_steps=steps, true_cfg_scale=cfg_scale, width=width, height=height, num_images_per_prompt=1, generator=generator, ) images.append((result.images[0], current_seed)) current_seed += 100 # separación segura return images @spaces.GPU(duration=70) def run_lora_multi( prompt_1, prompt_2, prompt_3, prompt_4, negative_prompt, cfg_scale, steps, selected_index, randomize_seed, seed, width, height, lora_scale, speed_mode, quality_multiplier, quantity, # se ignora acá (pero ahora lo usamos bien) history, prompt_enhance=False, progress=gr.Progress(track_tqdm=True) ): if selected_index is None: raise gr.Error("You must select a LoRA before proceeding.") prompts = [ p.strip() for p in [prompt_1, prompt_2, prompt_3, prompt_4] if p and p.strip() ] if not prompts: raise gr.Error("You must fill at least one prompt.") selected_lora = loras[selected_index] lora_path = selected_lora["repo"] trigger_word = selected_lora["trigger_word"] # aplicar trigger word por prompt final_prompts = [] for p in prompts: if trigger_word: if selected_lora.get("trigger_position") == "append": final_prompts.append(f"{p} {trigger_word}") else: final_prompts.append(f"{trigger_word} {p}") else: final_prompts.append(p) # limpiar LoRAs previas pipe.unload_lora_weights() # 🔥 CARGA DE LORAs (UNA SOLA VEZ) if speed_mode == "light 4": pipe.load_lora_weights( LIGHTNING_LORA_REPO, weight_name=LIGHTNING_LORA_WEIGHT, adapter_name="lightning" ) pipe.load_lora_weights( lora_path, weight_name=selected_lora.get("weights"), adapter_name="style" ) pipe.set_adapters(["lightning", "style"], adapter_weights=[1.0, lora_scale]) elif speed_mode == "light 8": pipe.load_lora_weights( LIGHTNING_LORA_REPO, weight_name=LIGHTNING8_LORA_WEIGHT, adapter_name="lightning" ) pipe.load_lora_weights( lora_path, weight_name=selected_lora.get("weights"), adapter_name="style" ) pipe.set_adapters(["lightning", "style"], adapter_weights=[1.0, lora_scale]) elif speed_mode == "Wuli-art": with calculateDuration("Loading Lightning LoRA and style LoRA"): pipe.load_lora_weights( 'Wuli-Art/Qwen-Image-2512-Turbo-LoRA', weight_name='Wuli-Qwen-Image-2512-Turbo-LoRA-4steps-V2.0-bf16.safetensors', adapter_name="lightning" ) weight_name = selected_lora.get("weights", None) pipe.load_lora_weights( lora_path, weight_name=weight_name, low_cpu_mem_usage=True, adapter_name="style" ) pipe.set_adapters(["lightning", "style"], adapter_weights=[1.0, lora_scale]) elif speed_mode == "light 4 fp8": with calculateDuration("Loading Lightning LoRA and style LoRA"): pipe.load_lora_weights( LIGHTNING_LORA_REPO, weight_name=LIGHTNING_FP8_4STEPS_LORA_WEIGHT, adapter_name="lightning" ) weight_name = selected_lora.get("weights", None) pipe.load_lora_weights( lora_path, weight_name=weight_name, low_cpu_mem_usage=True, adapter_name="style" ) pipe.set_adapters(["lightning", "style"], adapter_weights=[1.0, lora_scale]) else: pipe.load_lora_weights( lora_path, weight_name=selected_lora.get("weights"), adapter_name="style" ) pipe.set_adapters(["style"], adapter_weights=[lora_scale]) if randomize_seed: seed = random.randint(0, MAX_SEED) multiplier = float(quality_multiplier.replace("x", "")) width = int(width * multiplier) height = int(height * multiplier) # ✅ FIX: quantity viene como index 0..3 (por type="index"), convertimos a 1..4 real_quantity = int(quantity) + 1 if (history is None): history = [] gallery_images = [] for prompt in prompts: current_seed = seed if prompt_enhance: prompt = rewrite(prompt) # ✅ FIX: quantity ya no es el componente global; es un int real for _ in range(real_quantity): generator = torch.Generator(device="cuda").manual_seed(current_seed) result = pipe( prompt=prompt, negative_prompt=negative_prompt, num_inference_steps=steps, true_cfg_scale=cfg_scale, width=width, height=height, num_images_per_prompt=1, generator=generator, ) img = result.images[0] imgtuple = (img, str(current_seed)) # images.append(imgtuple) gallery_images.append(imgtuple) # history persistente (acumula) history = [(img, str(current_seed))] + history history = history[:24] yield gallery_images, history, history, seed current_seed += 100 # separación segura #return images #images = generate_images_for_prompts( # prompts=final_prompts, # negative_prompt=negative_prompt, # steps=steps, # seed=seed, # cfg_scale=cfg_scale, # width=width, # height=height, # quantity=real_quantity, # ✅ FIX: ahora se pasa # prompt_enhance=prompt_enhance, #) #gallery_images = [(img, str(s)) for img, s in images] #return gallery_images, seed # ... (El resto de las funciones como get_huggingface_safetensors, check_custom_model, etc., permanecen sin cambios) ... def get_huggingface_safetensors(link): split_link = link.split("/") if len(split_link) != 2: raise Exception("Invalid Hugging Face repository link format.") print(f"Repository attempted: {split_link}") model_card = ModelCard.load(link) base_model = model_card.data.get("base_model") print(f"Base model: {base_model}") acceptable_models = { "Qwen/Qwen-Image", "Qwen/Qwen-Image-2512", } models_to_check = base_model if isinstance(base_model, list) else [base_model] if not any(model in acceptable_models for model in models_to_check): raise Exception("Not a Qwen-Image LoRA!") image_path = model_card.data.get("widget", [{}])[0].get("output", {}).get("url", None) trigger_word = model_card.data.get("instance_prompt", "") image_url = f"https://huggingface.co/{link}/resolve/main/{image_path}" if image_path else None fs = HfFileSystem() try: list_of_files = fs.ls(link, detail=False) safetensors_name = None for file in list_of_files: filename = file.split("/")[-1] if filename.endswith(".safetensors"): safetensors_name = filename break if not safetensors_name: raise Exception("No valid *.safetensors file found in the repository.") except Exception as e: print(e) raise Exception("You didn't include a valid Hugging Face repository with a *.safetensors LoRA") return split_link[1], link, safetensors_name, trigger_word, image_url def check_custom_model(link): print(f"Checking a custom model on: {link}") if link.endswith('.safetensors'): if 'huggingface.co' in link: parts = link.split('/') try: hf_index = parts.index('huggingface.co') username = parts[hf_index + 1] repo_name = parts[hf_index + 2] repo = f"{username}/{repo_name}" safetensors_name = parts[-1] try: model_card = ModelCard.load(repo) trigger_word = model_card.data.get("instance_prompt", "") image_path = model_card.data.get("widget", [{}])[0].get("output", {}).get("url", None) image_url = f"https://huggingface.co/{repo}/resolve/main/{image_path}" if image_path else None except: trigger_word = "" image_url = None return repo_name, repo, safetensors_name, trigger_word, image_url except: raise Exception("Invalid safetensors URL format") if link.startswith("https://"): if link.startswith("https://huggingface.co") or link.startswith("https://www.huggingface.co"): link_split = link.split("huggingface.co/") return get_huggingface_safetensors(link_split[1]) else: return get_huggingface_safetensors(link) def add_custom_lora(custom_lora): global loras if custom_lora: try: title, repo, path, trigger_word, image = check_custom_model(custom_lora) print(f"Loaded custom LoRA: {repo}") model_card_examples = "" try: model_card = ModelCard.load(repo) widget_data = model_card.data.get("widget", []) if widget_data and len(widget_data) > 0: examples_html = '
{caption[:30]}{'...' if len(caption) > 30 else ''}
"+trigger_word+" as the trigger word" if trigger_word else "No trigger word found. If there's a trigger word, include it in your prompt"}